diff --git a/unsloth/models/gemma.py b/unsloth/models/gemma.py
index 646aec488d..e73d0a6688 100644
--- a/unsloth/models/gemma.py
+++ b/unsloth/models/gemma.py
@@ -518,8 +518,8 @@ class FastGemmaModel(FastLlamaModel):
         # Inferene can now be CUDAGraphed, but we shall retain the old rotary embeddings.
         # https://github.com/huggingface/transformers/pull/27931
         # https://github.com/huggingface/transformers/blob/v4.37.2/src/transformers/models/llama/modeling_llama.py
-        import transformers.models.gemma.modeling_gemma
-        transformers.models.gemma.modeling_gemma.GemmaRotaryEmbedding = LlamaRotaryEmbedding
+        # import transformers.models.gemma.modeling_gemma
+        # transformers.models.gemma.modeling_gemma.GemmaRotaryEmbedding = LlamaRotaryEmbedding
         return
     pass
 
diff --git a/unsloth/models/llama.py b/unsloth/models/llama.py
index 94fc954875..ecc09b2145 100644
--- a/unsloth/models/llama.py
+++ b/unsloth/models/llama.py
@@ -812,15 +812,11 @@ class LlamaRotaryEmbedding(torch.nn.Module):
         inv_freq = 1.0 / (
             self.base ** (torch.arange(0, self.dim, 2, dtype=torch.int64, device="cpu").float() / self.dim)
         )
-        position_ids = torch.arange(self.max_seq_len_cached, device="cpu", dtype=torch.int64).unsqueeze(-1)#.float()
-        inv_freq_expanded = inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
-        position_ids_expanded = position_ids[:, None, :].float()
-        freqs = (inv_freq_expanded @ position_ids_expanded).transpose(1, 2)
-        emb = torch.cat((freqs, freqs), dim=-1)
+        t = torch.arange(self.max_seq_len_cached, device="cpu", dtype=torch.int64).float()
 
-        # freqs = torch.outer(t, inv_freq)
+        freqs = torch.outer(t, inv_freq)
         # Different from paper, but it uses a different permutation in order to obtain the same calculation
-        # emb = torch.cat((freqs, freqs), dim=-1)
+        emb = torch.cat((freqs, freqs), dim=-1)
         self.register_buffer("cos_cached", emb.cos().to(dtype=dtype, device=device, non_blocking=True), persistent=False)
         self.register_buffer("sin_cached", emb.sin().to(dtype=dtype, device=device, non_blocking=True), persistent=False)
     pass